browser-harness is a thin, editable browser automation harness that connects an LLM directly to a real Chrome or Chromium browser via a single CDP websocket. During execution, the agent writes any missing helper code it needs, allowing it to complete arbitrary browser tasks.
Project overview
The harness improves itself on every run by automatically writing missing helper code and domain-specific skills during execution.
Project type
AI Agent
Use cases
Automation
Deployment
Refer to project documentation
License
MIT
Best for
Developers and AI engineers who need a thin, editable harness that lets an LLM drive a real browser and write its own helper code to complete arbitrary tasks.
Key capabilities
A thin, editable harness that connects an LLM directly to a real Chrome or Chromium browser via a single CDP websocket.
The agent writes missing helper code itself during execution if a required helper is missing, allowing it to complete arbitrary tasks.
Includes examples of reusable, site-specific domain skills (e.g., GitHub, LinkedIn, Amazon) that teach the agent selectors, flows, and edge cases.
The harness automatically writes domain skills during task execution when it figures out non-obvious flows, rather than requiring users to hand-author them.
Limitations and risks
Requires Chrome 144+ for per-attach remote debugging popup.
Requires a capable external LLM or coding agent to operate.
Getting started
Install or upgrade browser-harness using uv and Python 3.12. Run 'browser-harness skill' to register the skill. Open chrome://inspect/#remote-debugging and enable it. Then connect your coding agent to your browser. Setup difficulty is medium because it requires an external LLM or coding agent to paste and execute the setup prompt, and it requires configuring Chrome remote debugging.